Lake Huron's 1920s drop: three ways to detect it, one of them right

A strongly autocorrelated lake-level series through the changepoint detector: the plain model finds ten changes, AR(1) none, mean-and-variance the drop.

Tools and projects · 1 min read · Source
changepoint-detectortime-seriesmeasurement

changepoint-detector had been tested on level steps (the Nile), volatility breaks, counts and seasonal series, but never on a series where each year is strongly tied to the one before. Lake Huron's annual level, 1875 to 1972, is that: lag-1 autocorrelation of about 0.8, and a drop near 1920 in the literature.

modechanges found
plain mean10 (1889, 1923, 1929, 1931, 1942, 1951, 1956, 1963, 1966, 1969)
--ar1 (plain or joint)none
meanvar1889 and 1923 (range 1919 to 1925)

The plain model reads persistence as a string of small steps; the AR(1) model widens the noise so far that it also hides the drop. Only meanvar lands near 1920, and with one series that may be luck. The result is pinned in a test so a later fix has to show it moves these numbers. The open item is an AR(1) mean model that keeps a drop of this size.

The series is R's LakeHuron (98 numbers, source noted in the venture's data/SOURCE.md).

More in Tools and projects

All writing · All in this topic · All tools and research